How AI picks products.

An assistant asked to recommend a product doesn’t run a search and read the results. It assembles an answer from feeds, forums, roundups and its own priors — and most brands have no idea what it’s reading about them.

That gap is the whole problem. A head of ecommerce can tell you their organic rankings, their ROAS by channel, their Amazon buy-box rate. Almost none of them can tell you what ChatGPT says when a customer asks which product in their category is best. That answer is now a distribution channel, and it is being written by systems nobody in the building has instrumented.

This piece is about how those systems actually work: what they read, why traditional search performance doesn’t transfer, and what the emerging checkout layer changes. The landscape is moving fast enough that some of this will be stale within two quarters. I’ve flagged where the evidence is thin.

The behavior moved first

Adobe Analytics, which measures over a trillion visits to U.S. retail sites, found that traffic to U.S. retail sites from generative AI sources rose 1,200% in February 2025 compared with July 2024 — roughly doubling every two months since September 2024.[] Over the 2025 holiday season, retail traffic from generative AI tools grew 693.4% year over year, with Cyber Monday up 670%. Adobe was careful to note that “the base of users remains modest.”[]

That caveat matters more than the growth rate. Bain reported in November 2025 that AI accounts for up to 25% of referral traffic for some retailers, but less than 1% of total retail traffic overall.[] Anyone selling you an AI-visibility program on the strength of a four-figure percentage is selling you a derivative of a very small number.

The behavioral signal is more interesting than the traffic signal. Bain’s August 2025 analysis, built on Sensor Tower’s opt-in mobile panel, found that shopping queries rose from 7.8% to 9.8% of all ChatGPT prompts in the first half of 2025, and that click-through rates from ChatGPT nearly tripled between March and June, from 2.2% to 5.7%.[] A Clutch survey published in January 2026 found 65% of consumers use AI to research products before buying.[]

So the volume is small and the intent is dense. Which is exactly the profile of a channel worth instrumenting before it matters, rather than after.

Google rankings do not transfer

This is the finding most ecommerce teams have not absorbed.

In August 2025, Ahrefs ran 15,000 long-tail queries through Google and Bing, then put the same prompts to ChatGPT, Gemini, Copilot and Perplexity, and compared the URLs each cited against the search rankings. Only 12% of the links cited by ChatGPT, Gemini and Copilot appeared in Google’s top 10 for the same query. Roughly 80% of ChatGPT and Gemini citations didn’t rank anywhere in Google for that query at all. Perplexity was the outlier — close to one in three of its citations ranked in Google’s top 10.[]

Read that as an architectural statement, not a scoreboard. Assistants are not ranking pages. They are retrieving evidence, and the evidence they find most useful for a comparison question is frequently not the page a search engine would rank first for the same string.

Shopping is a partial exception, and the exception is instructive. Semrush ran 100 product prompts five times each in January 2026 and traced the hidden queries ChatGPT issues behind the scenes; the top ChatGPT product appeared in Google Shopping’s first three results 75% of the time. The authors flagged their own small sample.[] The pattern makes sense: for a product, the assistant needs a live catalog with price and availability, and that catalog is not the open web. For the reasoning around the product — which one is best, for whom, and why — it goes elsewhere.

The other thing rankings don’t prepare you for is how few slots exist. Semrush found that in consumer electronics, AI mentions an average of 1.22 brands per ChatGPT query and 1.36 per Google AI Mode query — typically one or two brands, not ten.[] Its expanded 2026 index, covering 126 million U.S. prompts from January to April 2026, found the top three consumer electronics brands accounted for 76.9% of total visibility in the category.[]

Page one had ten organic results. The answer has one or two names. That compression is the commercial fact of AI search.

Four inputs, weighted differently by platform

It helps to stop thinking about “the algorithm” and start thinking about four distinct input streams.

Structured commerce data. OpenAI’s own documentation states that ChatGPT considers “structured metadata from first-party and third-party providers (e.g., price, product description)” alongside the model’s own pre-search responses, and that “product results are selected independently by ChatGPT and are not ads, nor influenced by any OpenAI partnerships.”[] Merchants can apply to submit a direct feed; Shopify merchants are covered automatically via Shopify Catalog. The feed spec requires item ID, title (150 characters), description (up to 5,000), URL, brand, image, price, availability, seller details, a valid GTIN or MPN, and explicit is_eligible_search and is_eligible_checkout flags.[] Search Engine Land’s John Morabito put the distinction well in October 2025: in this model the feed is primary, not secondary — the trusted dataset the assistant reasons over rather than a supplement to crawling.[]

Google’s equivalent is older and vastly larger. Its Shopping Graph holds more than 50 billion product listings, two billion of which refresh every hour.[] Google’s guidance is explicit that on-page structured data and a Merchant Center feed are not alternatives: providing both “maximizes your eligibility to experiences,” and some surfaces combine the two.[]

Most retailers are not ready for this. Adobe’s April 2026 analysis scored U.S. retail pages for machine readability and found product detail pages the weakest surface, averaging 66 out of 100 — below homepages at 75 and category pages at 74.[] The page that has to carry the product is the page least legible to the machine reading it.

Editorial and affiliate roundups. XFunnel tracked 768,000 citations across ChatGPT, Google AI Overviews and Perplexity over twelve weeks, reported in April 2025. Product-related content made up 46–70% of citations depending on funnel stage, exceeding 70% at the decision stage. The consumer breakdown differs sharply from B2B: in B2C, product pages fell to roughly 35% of citations, while affiliate content rose to 18% and user reviews to 15%.[] For a consumer brand, that means well over a third of the evidence base is written by someone else.

Communities and reference. Profound tracked 680 million citations from August 2024 to June 2025. Wikipedia was ChatGPT’s single most-cited domain at 7.8% of all citations; Reddit led Google AI Overviews at 2.2% overall and made up roughly 47% of Perplexity’s top ten sources.[] A May 2026 synthesis by 5W Public Relations across nine independent datasets put Wikipedia at 13.15% and Reddit at 11.97% of U.S. ChatGPT citations — together over a quarter of the total — with the Wall Street Journal, New York Times, Bloomberg and Financial Times absent from the top twenty.[] These shares are volatile; Semrush observed Reddit’s ChatGPT citation frequency swinging hard within a single quarter. Treat the direction as durable and the specific number as perishable.

The model’s own priors. OpenAI says plainly that ChatGPT weighs “model responses generated by ChatGPT before it considers any new search results.”[] Some portion of what an assistant thinks about your category was fixed at training time and is not addressable by anything you publish this quarter.

How these four are weighted varies by platform, and the variance is large. Semrush found ChatGPT averaging around 15 sources per response against Gemini’s three.[] A source strategy tuned for one is not a strategy for the other.

The comparison query decides the revenue

Not all AI shopping queries are worth the same. Adobe’s survey of more than 5,000 U.S. consumers found the dominant uses were research (55%) and product recommendations (47%).[] That is the top and middle of the funnel collapsing into a single conversational turn.

The queries that matter are comparison queries: best running shoe for flat feet under $150, X vs Y for a small apartment, what should I buy instead of Z. They carry purchase intent, they name a shortlist, and the shortlist is one or two brands long.

Two mechanics make them different from keyword search. First, length: Google reported that AI Mode users write queries two to three times longer than in traditional search, on a surface that had passed 75 million daily active users by February 2026.[] Longer queries carry constraints — budget, use case, body type, climate — that a keyword never carried.

Second, decomposition. Google calls it query fan-out: the system runs multiple searches at once to work out the underlying criteria, its own example being what actually makes a bag suitable for rainy Portland weather, then matches products against those criteria.[] Semrush observed the same shape from the other side, decoding the hidden shopping queries ChatGPT generates from a natural-language prompt.[]

The practical consequence: you are not competing for a phrase. You are competing to be the obvious answer to a constraint the assistant invented on the fly. That is won by having attributes present in your feed, claims stated plainly enough to be extracted, and third-party corroboration for the specific comparison — not by a page optimized for a head term.

The agentic layer, and why it’s still unsettled

The recommendation layer is now growing a transaction layer beneath it.

Stripe and OpenAI launched Instant Checkout in ChatGPT on 29 September 2025, opening with U.S. Etsy sellers and over a million Shopify merchants to follow, and open-sourced the Agentic Commerce Protocol built around a Shared Payment Token scoped to a specific merchant and cart total.[] Google announced agentic checkout in November 2025 with Wayfair, Chewy, Quince and select Shopify stores, plus a Duplex-based agent that phones local stores to confirm stock.[] In May 2026 it introduced Universal Cart across Search and the Gemini app, extended the Universal Commerce Protocol internationally, and pushed its Agent Payments Protocol with merchants including Nike, Sephora, Target, Ulta Beauty, Walmart and Wayfair.[]

Two things temper this. Consumers are not there yet: Bain found roughly half of consumers uncomfortable with fully autonomous end-to-end AI transactions, and that they trust a retailer’s own agent about three times more than a third-party one.[] Clutch’s January 2026 survey was starker — only 4% would let AI complete a transaction on its own.[]

And the ground rules are genuinely unresolved. In March 2026, a federal judge in San Francisco granted Amazon a preliminary injunction blocking Perplexity’s Comet browser from operating inside logged-in Amazon accounts, finding the agent acted with the user’s permission but without Amazon’s authorization.[] On 4 August 2026, the Ninth Circuit overturned that injunction, holding Amazon hadn’t established a likely Computer Fraud and Abuse Act violation because users, not Perplexity, initiate the access. The case is still live.[]

Whether a retailer can refuse to serve an agent acting for its own customer is not a settled question. Build accordingly.

What this actually asks of you

Four things, roughly in order of payback.

Get the catalog machine-readable and keep it fresh — feed and on-page structured data, both, with real GTINs, accurate availability, and attributes that answer constraints rather than describe features. This is unglamorous plumbing and it is the part fully within your control.

Audit the third-party evidence base. For your top twenty comparison queries, find out what the assistants are actually citing. If it’s a roundup you’ve never contacted and a Reddit thread from 2023, that is your competitive position, whatever your rankings say.

Instrument the answer, not just the click. Semrush’s 2026 index found 45% of marketing leaders unable to measure brand visibility in AI answers, and only 9% with tools covering all relevant metrics across platforms.[] Last-click attribution will show you almost none of this.

Decide your agentic posture deliberately — which protocols you support, which agents you serve, whether your own on-site agent should be the trusted one. Bain’s trust data suggests that last question is worth more than it looks.

One honest note on direction. In March 2025, Adobe found AI-referred traffic converted 9% worse than other sources, down from a 43% gap the previous July.[] By March 2026 it had flipped: AI traffic converted 42% better than non-AI sources, against 38% worse a year earlier.[] A channel that inverted its conversion profile in twelve months is not a channel anyone should claim to have figured out. It is a channel worth being early in, measured carefully, with the plumbing already done.